Joint-MVSNet: Geometry-Controlled 2-D/3-D Interaction for Aerial Multiview Stereo Reconstruction

Abstract

Learning-based aerial multiview stereo (MVS) has advanced large-scale 3-D reconstruction from remote-sensing imagery; however, reliable depth inference remains difficult in urban scenes with large relief variation, limited view overlap, occlusion, weak texture, repeated patterns, and radiometric inconsistency. These factors corrupt cross-view sampling and produce broad or multimodal cost responses that can be amplified by cascade refinement. To address this problem, we propose Joint-MVSNet, a geometry-controlled framework that couples sparse 3-D priors with depthwise cost calibration and reliability-aware residual correction. First, the geometry message-controlled injection module backprojects coarse depth and posterior probability into sparse pseudopoints and writes reliability-weighted multiview geometry messages into the reference feature before cost-volume construction. Second, depthwise wavelet calibration decomposes depthwise cost sequences into low-frequency trends and high-frequency details and adaptively reweights these components to suppress ambiguous responses. Third, boundary prototype residual correction extracts structural prototypes from reliable boundary-sensitive regions and performs bounded corrections in low-confidence areas. Experiments yield mean absolute errors of 0.078 m on WHU-MVS, 0.108 m on LuoJia-MVS, and 0.1078 m on WHU-OMVS under their respective five-view evaluation settings.